{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/towards-an-understanding-of-neural-networks","title":"Towards an Understanding of Neural Networks in Natural-Image Spaces","arxiv_id":"1801.09097","date":"2018-01-27","proceeding":null,"authors":["Yifei Fan","Anthony Yezzi"],"abstract":"Two major uncertainties, dataset bias and adversarial examples, prevail in\nstate-of-the-art AI algorithms with deep neural networks. In this paper, we\npresent an intuitive explanation for these issues as well as an interpretation\nof the performance of deep networks in a natural-image space. The explanation\nconsists of two parts: the philosophy of neural networks and a hypothetical\nmodel of natural-image spaces. Following the explanation, we 1) demonstrate\nthat the values of training samples differ, 2) provide incremental boost to the\naccuracy of a CIFAR-10 classifier by introducing an additional \"random-noise\"\ncategory during training, 3) alleviate over-fitting thereby enhancing the\nrobustness against adversarial examples by detecting and excluding illusive\ntraining samples that are consistently misclassified. Our overall contribution\nis therefore twofold. First, while most existing algorithms treat data equally\nand have a strong appetite for more data, we demonstrate in contrast that an\nindividual datum can sometimes have disproportionate and counterproductive\ninfluence and that it is not always better to train neural networks with more\ndata. Next, we consider more thoughtful strategies by taking into account the\ngeometric and topological properties of natural-image spaces to which deep\nnetworks are applied.","url_abs":"http://arxiv.org/abs/1801.09097v2","url_pdf":"http://arxiv.org/pdf/1801.09097v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"towards-an-understanding-of-neural-networks","repo_url":"https://github.com/thelittlekid/natural-image-spaces","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"none","reach":null}],"tasks":[{"task_slug":"philosophy","task_name":"Philosophy"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}